Automatic Determination of Invariance in Machine Coded Speech

摘要

A method is described for determining significant patterns for the identification of spoken words automatically. The method is based on work of Gazdag, and uses his concept of significant subsequences (SSS, sequences of common machine events) that employs a machine representation of spoken words. An algorithm was developed and is described for computing the SSSs. A program was implemented on the ILLIAC II. Examples of the zeroth order of the program and the computed first and second order SSSs are given. In all computations applied to spoken digits, the SSSs converged at most by the second order SSS. Discussions of the data input and output and the flow chart of the program are appended. Details of the sampler developed for Gazdag's processor are given. Results suggest that the algorithm and sampler will enable identification of true SSSs for words other than spoken digits. (Author)